1

Click here to load reader

Segmentation and tracking algorithms on parallel hardware · Segmentation and tracking algorithms on parallel hardware ... Rensselaer Polytechnic Institute, ... and test various segmentation

Embed Size (px)

Citation preview

Page 1: Segmentation and tracking algorithms on parallel hardware · Segmentation and tracking algorithms on parallel hardware ... Rensselaer Polytechnic Institute, ... and test various segmentation

Segmentation and tracking algorithms on parallel hardware

SignificanceSegmenting blood vessels tagged by EBA is an important element of neuroscience and stem cell research. Segmentation of cell nuclei and tracking is of huge importance in immunology and cancer research. Challenge

• Low contrast, broken signal, huge 3D images => difficult to manually validate

• Huge memory requirement a challenge in a GPU implementation.

• Achieving cell segmentation and tracking in a BlueGene supercomputer has issues with both memory and requires a highly data parallel design for high speedups.

State-of-the-Art• Robust 3-D Modeling of Vasculature Imagery Using Superellipsoids

• Spatio-temporal cell cycle phase analysis using level sets and fast marching methods

"This work was supported in part by Gordon-CenSSIS, the Bernard M. Gordon Center for Subsurface Sensing and Imaging Systems, under the Engineering Research Centers Program of the National ScienceFoundation (Award Number EEC-9986821)."

Technical ApproachVessel segmentation on GPU: 3-D data sets containing separate five data channels in 100μm thick tissue slices was collected using spectral unmixingtechniques. Typical dataset size: 1024x1024x60.

Generalized likelihood ratio based hypothesis testing was used for vessel detection step.

Contact info.Badrinath Roysam , ProfessorDept. of Electrical, Computer, and Systems EngineeringRensselaer Polytechnic Institute, 110 8th Street, Troy, NY 12180Phone: (518)276-8067; Fax: 518-276-8715;Email: [email protected]

ReferencesRobust Adaptive 3-D Segmentation of Vessel Laminae from Fluorescence Confocal Microscope Images & Parallel GPU Implementation. Arunachalam Narayanaswamy, Saritha Dwarakapuram, Christopher S. Bjornsson, Barbara M. Cutler, William Shain, Badrinath Roysam,, (accepted, in press), IEEE Transactions on Medical Imaging, March 2009.

Spatio-temporal cell cycle phase analysis using level sets and fast marching methods. Padfield D, Rittscher J, Thomas N, Roysam B. Med Image Anal. Jul 23, 2008 .

Arunachalam Narayanaswamy*, Saritha Dwarakapuram*, Chris Carothers¤, Barbera Cutler¤, Chris Bjornsson§,William Shain¶, Ellen Robey� and Badrinath Roysam*

Department of Electrical Engineering*, Computer Science ¤ and Biomedical Engineering§ , Rensselaer Polytechnic Institute, Troy

Center for Neural Communication Technology ¶, Wadsworth Center, Albany

Department of Molecular and Cell Biology�, University of California, Berkeley

AbstractWe present a GPU implementation of the Vessel Surface segmentation algorithm we developed using robust hypothesis testing and an adaptive region growing algorithm. We also extract the centerlines from the surface segmentation. We provide a comparison of our results with the existing Hessian filtering based technique and achieved an 8X speed up with the GPU implementation. We also present our work on Cell segmentation and tracking that runs on the BlueGene Supercomputer at RPI. This achieves a speedup of upto 600x to unmix, segment and track a series of 764 3-D images. This greatly reduces the feedback time enabling us to try out new algorithms to segment and track cells.

(A) The two hypotheses (B) maximum-intensity projection of the original 3-D confocaldataset that is contrast stretched to better show the imaging noise and overlap from the nuclear channel (small circular objects). (C) Adaptive Otsu thresholding; (D) Hessian filtering (E) Tube tracing results [4]; (F) Basic surface extraction algorithm described in Section D; (G) Results of adaptive contiguous extension (Section E). These 2 panels have the same sensitivity (true positive rate).

Illustrating the parallel implementation in GPU (A) A table showing the performance results in CPU/GPU (B) A flowchart of the surface segmentation algorithm with ACE. We implemented 5 different kernel functions to achieve a parallel implementation in the GPU.

Centerline extraction by voting and tube tracing

Cell Segmentation and Tracking(Left) A 4 channel image showing the two types of T cells, dendritic cells and blood vessel at one time point of a time lapse data. (Right) A 3-D rendering of the segmented and tracked cells for the same time point.

We developed a parallel implementation of the cell segmentation and tracking using MPI libraries and tested it in the BlueGene Supercomputer at RPI. The segmentation and tracking was done over 764 3-D image stacks and the speedup achieved was over 600X from a serial computer. The work is in progress and we hope to develop and test various segmentation and tracking algorithms much faster since the computational time for such a large dataset is less than 2 minutes.

t1 t2 t1 t2 …

Send feature information to the previous time point

Do pair wise tracking in parallel

Send IDs

Send features

Dataset 1 Dataset 2 Dataset n

Parallel tracking algorithm

1

3 4

2

Time point 1

1

3 4

2

Time point 2

1

3 4

2

Time point 3

1

3 4

2

Time point n

…….

1

3 4

2

Time point 1

1

3 4

2

Time point 2

1

3 4

2

Time point 3

1

3 4

2

Time point n

…….

Preprocessing …….

Inte

ract

Inte

ract

Inte

ract

Inte

ract

Input

Output

Parallel preprocessing

Preprocessing is parallel across all time points, but is serial among

all channels at a single time point. Tracking has parallelism in various levels. Initial pair-wise correspondence between consecutive time points is computed in parallel. Then the indexes are transferred serially in a single dataset and parallel across different datasets.